bigscience-data

/

sgpt-bloom-1b7-nli

like 11

The command looks fine to me - did training already finish? If not, which error did you get?

Yes, if you have good French data available, I would expect slightly better performance by training on it.
You can try with the French STS datasets from the link you sent 👍

Let me know how it goes!

The training has finished: https://huggingface.co/Mayhem50/sgpt-bloom-560M-nli
But I was expecting better score on my dataset.

I will try to fine-tune both and see if improvements are significant

Thanks a lot.

Oh nice!
Note that the gap between BitFit & full fine-tuning only diminishes as you increase model size. For 560 million parameters you are likely better off training without BitFit (i.e. remove the --freezenonbias from your command).
If you scale up to 1.7B like this model or 7.1B like https://huggingface.co/bigscience/sgpt-bloom-7b1-msmarco, BitFit should perform just as well as full fine-tuning, so you can keep the command as is.

Also make sure that your downstream task is a symmetric one. If it's search-related, you may be better off training on MSMARCO.

I use this command to train, CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 accelerate launch /content/code/biencoder/nli_msmarco/sentence-transformers/examples/training/nli/training_nli_v2.py --model_name bigscience/bloom-560m --train_batch_size 64 --lr 32e-5 --pooling weightedmean --wandb --wandbwatchlog gradients --gradcache --chunksize 4

But it cannot be parallel. Using multiple GPUs is the same as one GPU. What's the problem?

Maybe you have to run accelerate config and select multiple gpus